HRGNN: hierarchical region-aware graph neural network for interpretable EEG-Based emotion recognition

Objective.Electroencephalography (EEG)-based emotion recognition has increasingly adopted graph neural networks (GNNs) to model functional connectivity. However, most existing approaches operate at the electrode level without explicitly incorporating functional brain region priors, leading to limited granularity in regional representation. Furthermore, conventional GNN architectures often suffer from over-smoothing, and global pooling strategies may obscure critical region-specific information. This study aims to develop a region-aware hierarchical graph framework that better aligns with neurophysiological organization while enhancing discriminative representation and interpretability.Approach.We propose a Hierarchical region-aware GNN (HRGNN). First, a brain region embedding module integrates anatomical partition priors to transform whole-brain EEG signals into structured regional subgraphs. A region-aware graph encoder is then employed to capture multi-scale intra- and inter-regional interactions through regional aggregation and hierarchical pooling. To further enhance discriminative power, a dynamic routing-based mixture-of-experts module adaptively fuses regional representations by assigning higher weights to emotionally salient brain regions.Main results.Extensive evaluations across eight publicly available EEG emotion recognition datasets demonstrate that HRGNN consistently outperforms state-of-the-art methods in both within-subject and cross-subject settings. The model achieves improved classification accuracy and robustness while mitigating over-smoothing effects. Visualization analyses reveal region-specific contribution patterns that are consistent with established neuroscientific findings on functional brain interactions during emotional processing.Significance.By integrating hierarchical region-aware modeling with adaptive expert fusion, HRGNN bridges graph-based learning and neurophysiological priors. The proposed framework improves performance, interpretability, and cross-subject generalization, offering a principled approach for EEG-based affective brain-computer interfaces and neuroengineering applications.